Single-Shot Object Detector Based on Attention Mechanism

For object detection, two-stage detectors have been achieving the dominant performance while one-stage detectors have the advantage of high efficiency. In order to better combine the advantages of both, we innovatively introduce the attention mechanism into single-shot object detector. Specifically, we introduce the IoU predictor to forecast the IoU between each predicted bounding box and corresponding ground-truth and propose an IoU-attention module to resample the feature map. Moreover, we introduce SE unit into our feature fusion module to encourage the model to focus on important information. We train the proposed model in an end-to-end way via multi-task loss function. With high efficiency, our model achieves competitive performance among start-of-the-art detectors on the challenging PASCAL VOC 2007, PASCAL VOC 2012 and MS COCO datasets.